[
https://issues.apache.org/jira/browse/CASSANDRA-15790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17102783#comment-17102783
]
Yifan Cai commented on CASSANDRA-15790:
---------------------------------------
Thanks for digging deeper into call sites for readValue.
I agree that the overrides in {{EmptyType}} should never be called with the
current usage. All the usage are guarded by empty value check and produce an
{{EMPTY_BYTE_BUFFER}} inline (without even calling the {{readValue()}} method),
except one. {{SinglePartitionReadCommand.Deserializer#deserialize}} does call
the {{readValue()}} method regardlessly. However, the type is for partition
keys, so I do not think it could be an EmptyType.
I think returning a value (either {{EMPTY_BYTE_BUFFER}} or null) from readValue
method sounds better. Because {{EmptyType}} cannot have the assumption on how
the call sites do the check and skip calling the readValue method.
- Returning null, eventually cause an exception to indicate the abnormal state.
- Returning {{EMPTY_BYTE_BUFFER}} may silent the error.
_I did not read carefully and thought {{maxValueSize}} was the length to read.
Thanks for pointing it out._
> EmptyType doesn't override writeValue so could attempt to write bytes when
> expected not to
> ------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-15790
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15790
> Project: Cassandra
> Issue Type: Bug
> Components: CQL/Semantics
> Reporter: David Capwell
> Assignee: David Capwell
> Priority: Normal
> Fix For: 3.0.x, 3.11.x, 4.0-alpha
>
>
> EmptyType.writeValues is defined here
> https://github.com/apache/cassandra/blob/e394dc0bb32f612a476269010930c617dd1ed3cb/src/java/org/apache/cassandra/db/marshal/AbstractType.java#L407-L414
> {code}
> public void writeValue(ByteBuffer value, DataOutputPlus out) throws
> IOException
> {
> assert value.hasRemaining();
> if (valueLengthIfFixed() >= 0)
> out.write(value);
> else
> ByteBufferUtil.writeWithVIntLength(value, out);
> }
> {code}
> This is fine when the value is empty as the write of empty no-ops (the
> readValue also noops since the length is 0), but if the value is not empty
> (possible during upgrades or random bugs) then this could silently cause
> corruption; ideally this should throw a exception if the ByteBuffer has data.
> This was called from
> org.apache.cassandra.db.rows.BufferCell.Serializer#serialize, here we check
> to see if data is present or not and update the flags. If data is present
> then and only then do we call type.writeValue (which requires bytes is not
> empty). The problem is that EmptyType never expects writes to happen, but it
> still writes them; and does not read them (since it says it is fixed length
> of 0, so does read(buffer, 0)).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
